Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method comprising: reading a neural network description describing a neural network comprising a plurality of functional units on a plurality of cores; selecting a functional unit of the plurality of functional units; dividing the functional unit into a plurality of subunits; connecting the plurality of subunits to the neural network in place of the functional unit; reallocating the plurality of functional units and the plurality of subunits between the plurality of cores; removing one or more unused core of the plurality of cores; writing an optimized neural network description based on the reallocation.
A method for optimizing the execution of a neural network on a multi-core processing system. The process begins by reading a neural network description that defines a neural network composed of multiple functional units distributed across several processor cores. A specific functional unit is selected for optimization, and this unit is divided into smaller subunits. These subunits are then integrated back into the neural network in place of the original functional unit. The system reallocates all functional units and newly created subunits across the available processor cores to improve efficiency. Any processor cores that are no longer required due to the reallocation are removed from the system configuration. Finally, an optimized neural network description is generated, reflecting the improved distribution of computational workload across the remaining cores. This approach aims to enhance the performance and resource utilization of neural network processing by dynamically restructuring the network and its execution environment.
2. The method of claim 1 , wherein the neural network description and the optimized neural network description have substantially the same functionality.
The invention relates to neural network optimization techniques, specifically ensuring that an optimized neural network retains the same functionality as the original neural network description. The method involves transforming a neural network description into an optimized version while maintaining identical behavior, meaning the optimized network produces the same outputs for all possible inputs as the original. This is achieved through techniques such as pruning, quantization, or architectural modifications that reduce computational complexity without altering the network's decision-making process. The optimization process may include validating the optimized network against the original to confirm functional equivalence, ensuring no loss of accuracy or performance. The approach addresses challenges in deploying neural networks on resource-constrained devices where efficiency is critical, but where maintaining model fidelity is essential. By preserving functionality, the optimized network can replace the original in applications such as edge computing, embedded systems, or cloud-based inference without requiring retraining or adjustments to downstream systems. The method may also involve intermediate representations or intermediate steps to facilitate the optimization while ensuring equivalence is maintained throughout the transformation.
3. The method of claim 1 , wherein the functional unit comprises a splitter.
The invention relates to a functional unit within a system designed to process or distribute signals, where the functional unit includes a splitter. The splitter divides an input signal into multiple output signals, enabling parallel processing or distribution of the signal across different pathways. This configuration is particularly useful in communication systems, signal processing applications, or network architectures where a single input needs to be replicated or shared among multiple components without degradation of signal integrity. The splitter ensures that the original signal is accurately divided, maintaining consistency in amplitude, phase, or other critical parameters across all outputs. This approach optimizes resource utilization by allowing a single source to serve multiple destinations simultaneously, reducing redundancy and improving efficiency in the overall system. The functional unit may be integrated into larger architectures, such as data transmission systems, audio/video distribution networks, or sensor arrays, where signal duplication is essential for scalability and performance. The splitter's role is foundational in enabling distributed processing or multi-point connectivity, making it a key component in systems requiring high reliability and low latency.
4. The method of claim 1 , wherein dividing the functional unit comprises: identifying at least one axon providing input to the functional unit; replicating the at least one axon; and connecting the replicated axon to one of the plurality of subunits.
The invention relates to a method for dividing a functional unit in a neural network or computational system into multiple subunits to improve processing efficiency or fault tolerance. The core problem addressed is how to distribute the input connections of a functional unit across its subunits while maintaining the original functionality. The method involves identifying at least one axon (input connection) that provides input to the functional unit. This axon is then replicated, creating an additional copy of the input connection. The replicated axon is subsequently connected to one of the multiple subunits within the divided functional unit. This process ensures that the input signal is distributed across the subunits, allowing parallel processing or redundancy. By replicating and redistributing axons, the method enables the functional unit to be split into subunits without losing input connectivity, which could be useful in scalable neural architectures or fault-tolerant computing systems. The replication step ensures that each subunit receives the necessary input signals, maintaining the integrity of the functional unit's operation.
5. The method of claim 1 , wherein connecting the plurality of subunits to the neural network comprises: adding a splitter to the neural network; and connecting the splitter to each of the plurality of subunits.
A neural network system and method for integrating multiple processing subunits into a neural network architecture. The invention addresses the challenge of efficiently connecting and managing multiple subunits within a neural network to enhance computational performance or modularity. The system involves adding a dedicated splitter component to the neural network, which acts as an intermediary between the neural network and the plurality of subunits. The splitter is configured to distribute input data from the neural network to each subunit and aggregate their outputs back into the network. This design allows for parallel processing across subunits while maintaining a unified interface with the neural network. The splitter may include mechanisms for load balancing, data routing, or synchronization to optimize the interaction between the neural network and the subunits. The subunits themselves could be specialized processing units, such as additional neural network layers, feature extractors, or other computational modules designed to augment the neural network's functionality. The splitter ensures seamless integration by handling the communication protocols, data formatting, and coordination required to incorporate these subunits without altering the core neural network structure. This approach enables scalable and flexible neural network architectures that can dynamically adapt to different computational tasks by leveraging the capabilities of multiple subunits.
6. The method of claim 1 , wherein connecting the plurality of subunits to the neural network comprises: connecting an axon of one of the plurality of subunits to a neuron of another of the plurality of subunits.
The invention relates to a neural network architecture designed to simulate biological neural connections, specifically addressing the challenge of efficiently linking multiple computational subunits to mimic synaptic interactions. The core problem involves enabling dynamic and biologically plausible communication between subunits to enhance the network's learning and processing capabilities. The method involves connecting an axon from one computational subunit to a neuron in another subunit, forming a direct synaptic link. This connection allows for the transmission of signals between subunits, enabling the network to process information in a manner analogous to biological neural systems. By establishing these axon-to-neuron connections, the network can propagate and integrate signals across subunits, facilitating more complex and adaptive learning behaviors. This approach contrasts with traditional neural networks that rely on fixed or predefined connections, as it introduces a biologically inspired mechanism for dynamic signal routing. The method supports modularity and scalability, allowing subunits to be added or reconfigured without disrupting the overall network structure. The axon-to-neuron connection ensures that signals are directed to specific targets, improving the precision and efficiency of information processing within the network.
7. The method of claim 1 , wherein the plurality of subunits and the reallocation are determined by application of an optimization model.
The invention relates to a method for managing a distributed system comprising multiple subunits, where resources or tasks are dynamically reallocated among the subunits to optimize performance. The core of the method involves determining both the configuration of the subunits and the reallocation process through the application of an optimization model. This model evaluates various factors such as current workload, resource availability, and system constraints to identify the most efficient distribution of tasks or resources. By leveraging this optimization-driven approach, the system can adapt in real-time to changing conditions, ensuring balanced utilization and improved overall efficiency. The method may incorporate algorithms that consider historical data, predictive analytics, or heuristic rules to guide the optimization process. The goal is to enhance system responsiveness, reduce bottlenecks, and maximize throughput without manual intervention. This approach is particularly applicable to large-scale computing environments, such as cloud computing platforms, distributed computing clusters, or multi-processor systems, where dynamic resource management is critical for maintaining performance.
8. The method of claim 7 , wherein the optimization model solves simultaneously for optimal division and reallocation.
The invention relates to a computational optimization system designed to improve resource allocation and partitioning in complex systems. The core problem addressed is the need for simultaneous determination of the most efficient division of a system into subsystems and the optimal redistribution of resources among those subsystems. Traditional approaches often handle these tasks sequentially, which can lead to suboptimal outcomes due to interdependencies between division and allocation decisions. The described method integrates both processes into a single optimization model, allowing for a unified solution that accounts for their mutual influence. This approach leverages mathematical programming or heuristic techniques to evaluate multiple configurations and resource distributions concurrently, ensuring that the final solution maximizes overall system efficiency or minimizes costs. The model may incorporate constraints such as capacity limits, cost thresholds, or performance requirements to ensure feasibility. By solving for both optimal division and reallocation in one step, the system avoids the inefficiencies of iterative or separate optimization processes, providing a more robust and coordinated solution. Potential applications include supply chain management, network design, manufacturing systems, or any domain requiring dynamic partitioning and resource distribution.
9. A computer program product for optimizing a neurosynaptic network, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to perform a method comprising: reading a neural network description describing a neural network comprising a plurality of functional units on a plurality of cores; selecting a functional unit of the plurality of functional units; dividing the functional unit into a plurality of subunits; connecting the plurality of subunits to the neural network in place of the functional unit; reallocating the plurality of functional units and the plurality of subunits between the plurality of cores; removing one or more unused core of the plurality of cores; writing an optimized neural network description based on the reallocation.
A computer program product for optimizing a neurosynaptic network, which is a type of artificial neural network designed to mimic the brain's structure and function. The invention addresses the challenge of efficiently distributing and reallocating computational resources in such networks to improve performance and reduce resource waste. The program operates on a neural network description that defines a network composed of multiple functional units distributed across several processing cores. It begins by selecting a specific functional unit for optimization. This unit is then divided into smaller subunits, which are reconnected to the network in place of the original unit. The system reallocates both the original functional units and the newly created subunits across the available cores, ensuring a more efficient distribution of computational load. Any cores that are no longer required due to this reallocation are removed from the system. Finally, the program generates an optimized neural network description that reflects these changes, resulting in a more streamlined and efficient network configuration.
10. The computer program product of claim 9 , wherein the neural network description and the optimized neural network description have substantially the same functionality.
The invention relates to neural network optimization in computer systems, specifically ensuring that an optimized neural network retains the same functional behavior as the original neural network. The technology involves a computer program product that processes a neural network description, applies optimization techniques to generate an optimized version, and verifies that both versions produce identical outputs for the same inputs. This addresses the problem of maintaining functional equivalence during neural network compression, pruning, or other optimization processes, which is critical for deploying efficient models without altering their decision-making capabilities. The solution includes mechanisms to compare the original and optimized networks, ensuring no loss of accuracy or unintended behavioral changes. By preserving functionality, the optimized network can be safely deployed in production environments where reliability is essential, such as in safety-critical applications like autonomous systems or medical diagnostics. The approach likely involves techniques such as layer-wise equivalence checks, output comparison across test datasets, or formal verification methods to validate that the optimized network behaves identically to the original under all relevant conditions.
11. The computer program product of claim 9 , wherein the functional unit comprises a splitter.
The invention relates to a computer program product designed to process data streams, where the product includes a functional unit responsible for dividing incoming data into separate components. This functional unit, referred to as a splitter, operates by receiving a continuous data stream and partitioning it into distinct segments or channels based on predefined criteria, such as time intervals, data types, or priority levels. The splitter ensures that the data is organized efficiently for subsequent processing stages, which may involve analysis, storage, or transmission. By separating the data into manageable parts, the splitter enhances the overall performance and scalability of the system, particularly in scenarios where high-volume or real-time data processing is required. The functional unit may also include additional components, such as filters or aggregators, to further refine the data before it is passed to other modules within the program. This approach improves throughput and reduces latency, making it suitable for applications in fields like telecommunications, financial services, or large-scale data analytics. The splitter's design allows for dynamic adjustment of partitioning rules, enabling adaptability to varying data characteristics or processing demands.
12. The computer program product of claim 9 , wherein dividing the functional unit comprises: identifying at least one axon providing input to the functional unit; replicating the at least one axon; and connecting the replicated axon to one of the plurality of subunits.
The invention relates to a computer program product designed to optimize the division of a functional unit within a neural network or similar computational model. The core problem addressed is the efficient partitioning of neural components to enhance parallel processing or distributed computation while maintaining functional integrity. The solution involves a method for dividing a functional unit by first identifying at least one axon that provides input to the functional unit. An axon is a neural structure that transmits signals between neurons or computational units. The method then replicates this identified axon, creating an additional pathway for signal transmission. Finally, the replicated axon is connected to one of multiple subunits within the functional unit, ensuring that the replicated pathway integrates seamlessly into the existing structure. This approach allows for the distribution of computational load across subunits, potentially improving performance in parallel processing environments. The replication and redistribution of axons help maintain the functional relationships within the network while enabling scalable and efficient division of the functional unit. The method is particularly applicable in neural network architectures where modularity and parallelism are critical for performance optimization.
13. The computer program product of claim 9 , wherein connecting the plurality of subunits to the neural network comprises: adding a splitter to the neural network; and connecting the splitter to each of the plurality of subunits.
A system for integrating multiple subunits into a neural network involves adding a splitter component to the neural network architecture. The splitter acts as an intermediary, linking each of the subunits to the neural network. This configuration allows the neural network to process inputs from or distribute outputs to the subunits efficiently. The splitter ensures that data flows between the subunits and the neural network are managed in a structured manner, potentially improving the overall performance and scalability of the system. The approach is designed to facilitate modular integration of subunits without requiring direct modifications to the neural network's core structure.
14. The computer program product of claim 9 , wherein connecting the plurality of subunits to the neural network comprises: connecting an axon of one of the plurality of subunits to a neuron of another of the plurality of subunits.
The invention relates to a computer program product designed to simulate or model neural networks using interconnected subunits, where each subunit represents a neuron or neural component. The core problem addressed is enabling dynamic and biologically plausible connections between these subunits to replicate neural network behavior more accurately. The program product connects multiple subunits by establishing links between their components, specifically by linking an axon of one subunit to a neuron of another subunit. This mimics the biological process of synaptic transmission, where an axon from one neuron connects to a dendrite or cell body of another neuron. The connection allows for signal propagation between subunits, enabling the neural network to process information in a manner analogous to biological neural systems. The program product likely includes additional features referenced in claim 9, such as mechanisms for initializing the neural network, defining subunit properties, or managing signal transmission rules. These features collectively support the creation of a functional neural network model where subunits interact through biologically inspired connections. The invention aims to provide a more realistic simulation of neural processes for applications in neuroscience research, artificial intelligence, or brain-machine interfaces.
15. The computer program product of claim 9 , wherein the plurality of subunits and the reallocation are determined by application of an optimization model.
This invention relates to optimizing the allocation of computational resources in a distributed computing environment. The problem addressed is the inefficient use of computing resources when tasks are not optimally distributed across available processing units, leading to bottlenecks, underutilization, or excessive energy consumption. The invention involves a computer program product that manages the distribution of computational workloads across multiple processing subunits. The program dynamically reallocates tasks among the subunits to improve performance, reduce latency, or minimize energy usage. The key innovation is the use of an optimization model to determine the optimal configuration of subunits and the reallocation of tasks. The optimization model evaluates factors such as processing capacity, task dependencies, and energy efficiency to generate an optimal distribution strategy. The reallocation process is automated, ensuring that tasks are reassigned in real-time based on changing conditions, such as workload fluctuations or hardware availability. This approach enhances system efficiency by balancing the load across subunits and adapting to dynamic computing environments. The optimization model may incorporate machine learning or mathematical programming techniques to refine the allocation strategy over time. The invention is particularly useful in cloud computing, high-performance computing, and edge computing systems where resource management is critical.
16. The computer program product of claim 15 , wherein the optimization model solves simultaneously for optimal division and reallocation.
The invention relates to a computer program product designed to optimize resource allocation and division within a system. The core functionality involves an optimization model that concurrently determines the most efficient way to divide a given set of resources and reallocate them to different tasks or entities. This simultaneous solving approach aims to maximize overall system efficiency, minimize waste, or achieve other predefined objectives by considering both the partitioning of resources and their redistribution in a single integrated process. The optimization model likely employs mathematical algorithms, such as linear programming, mixed-integer programming, or heuristic methods, to evaluate multiple possible configurations and select the optimal one based on specified constraints and goals. By addressing division and reallocation in tandem, the system avoids suboptimal solutions that might arise from treating these steps separately, such as inefficiencies in resource utilization or conflicting priorities between tasks. The computer program product could be applied in various domains, including supply chain management, workforce scheduling, financial portfolio optimization, or energy distribution, where dynamic and simultaneous adjustments to resource allocation are critical for performance. The innovation lies in the integrated approach to solving these interdependent problems, rather than addressing them sequentially or independently.
Unknown
April 28, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.